Agent Experience (AX): Designing for the AI Agents Acting on Behalf of Your Users

On June 3, 2026, Cloudflare CEO Matthew Prince posted on X that bots had passed human traffic on the web for the first time. Cloudflare Radar’s bot-vs-human view, filtered to HTML traffic, showed 57.5% of requests coming from bots versus 42.5% from humans. Prince had told an SXSW audience in March that the crossover would come in 2027, so it arrived well ahead of his own forecast.

This is a structural shift in how the web gets used, with direct implications for any team currently designing or redesigning a digital platform. Bot traffic and AI agent traffic aren’t the same thing, of course. Automated traffic includes traditional crawlers, monitoring systems, scrapers, and malicious bots. But the crossover points to a broader change: machines increasingly consume the web, and a growing share of them act on behalf of users instead of only indexing pages.

For years, digital design has revolved around two people: the end user and the developer. UX taught us that technology has to adapt to people, not the other way around. DX extended that principle to the people building the systems. If the API is confusing, the documentation is missing, or the development environment is frustrating, the product suffers before it even reaches the user.

Now a third actor is interacting with our platforms in a completely different way: the AI agent. An emerging discipline is taking shape around it, increasingly described as AX, or Agent Experience.

UX, DX, AX: Three perspectives, one product

UX asks how a person perceives and interacts with an interface. Its metrics are cognitive and emotional: clarity, simplicity, accessibility, time to task completion. Good UX means clear visual hierarchy, microcopy that doesn’t hide anything, intuitive flows, and accessibility that doesn’t block anyone.

DX asks a similar question about a different kind of user: the developer consuming an API, configuring an SDK, or integrating a service. What matters is consistent contracts, useful error messages, good documentation, and a fast feedback loop. Good DX lets developers think about their own problem instead of fighting the tool.

AX adds a third layer: how does an autonomous agent navigate, understand, and operate within our digital product? An agent doesn’t perceive a platform the way a person does, and it doesn’t consume it the way a traditional API integration does either. It acts, reading pages, interpreting content, weighing options, deciding, and executing steps, often iteratively and without anyone watching every move.

The new shape of web traffic

According to HUMAN Security’s 2026 State of AI Traffic & Cyberthreat Benchmark Report, AI agent traffic grew 7,851% year over year, and AI-driven traffic as a whole nearly tripled in volume throughout 2025. That number needs context. Agentic traffic started from a tiny base in 2024, so most of that huge percentage reflects how early the category was. Even so, HUMAN found that by the end of 2025 its absolute volume had grown large enough to register across major digital properties.

HUMAN separates agents from traditional crawlers and scrapers by what they do: navigating pages, filling forms, comparing products, starting transactions, and managing account workflows.

The asymmetry is easy to picture. When a human searches for a hotel in Barcelona for the weekend, they might check a handful of sites. When an agent does the same task, it can potentially inspect hundreds of pages, compare options, revisit results, and run through multiple steps before it hands the user an answer.

WorkOS documents this precisely: human traffic has not declined; what has happened is that each human action now generates orders of magnitude more machine requests. Increasingly, that activity comes from systems that behave nothing like traditional crawlers.

According to HUMAN Security’s State of Agentic Traffic report for April 2026, agentic browsers accounted for roughly 71% of observed activity among its top 10 agents, led by Perplexity’s Comet at 48.12% and OpenAI’s Atlas at 21.33%. The Claude Chrome extension followed at 17.33%. Unlike crawlers that build an index, agentic browsers arrive with user-agent strings, cookies, and session patterns that look a lot like human browsing.

The industry breakdown is just as telling. In April, media accounted for 45.62% of observed agentic traffic, e-commerce 38.20%, and travel 14.12%. Together those three categories made up roughly 98% of the activity HUMAN observed.

The landscape moves fast, too. By June, e-commerce had moved ahead of media, and Claude had overtaken Atlas as the second-largest source of observed agentic traffic. The exact rankings will keep shifting; the direction matters more than any single snapshot.

Agents are becoming participants in digital journeys well beyond browsing. HUMAN’s research shows agent activity reaching authentication, account, checkout, and payment flows, evidence that autonomous systems are moving past information retrieval and into real interactions.

That matters directly for retail, banking, insurance, bookings, travel, and service management, many of the same environments where Mimacom designs and builds digital platforms. The important question is whether our platforms are understandable and operable when machines act for real customers.

What an agent actually needs

An AI agent booking a flight, comparing insurance products, or managing a service request doesn’t read a page the way a human does. It ignores the image carousel and the visual hierarchy, and looks for semantic signals instead: what a button does, what submitting a form commits the user to, whether a price is final, what happens next, whether the state of the process has changed, and which conditions apply to an option.

When AX is poor, agents tend to stumble in a few predictable ways:

• Semantic ambiguity. Generic controls like “Continue,” “Next,” or “Submit” make sense visually to a human who can read the surrounding interface, but they don’t always say enough about what the action actually does. An agent shouldn’t have to guess what happens after a click.
• Flow fragmentation. Critical information often gets scattered across modals, tooltips, dropdowns, tabs, or elements that only appear after a prior interaction. A human can explore that visually, but an agent may never find it, or find it too late.
• Non-parseable structures. Pricing comparisons, product specifications, and contract conditions sometimes live inside images or visual layouts with no real semantic structure behind them. The information is technically visible but programmatically ambiguous.
• Hidden state. When a discount applies, a price changes, a booking moves to the next stage, or an application goes incomplete, none of it may show up clearly in the DOM, the URL, or another machine-readable state. The agent then has to infer what happened, and inference is where mistakes start.

Poor AX makes agents slower and, more importantly, less reliable. The agent acts on information it misread, missed, or assumed wrong, and the person it’s working for bears the consequences.

Why AX matters now, in the design of your next project

Filing AX under “something we’ll need eventually” would be a mistake. When a team starts designing an e-commerce platform, a financial services portal, a booking system, or a customer service experience, the question is no longer just whether customers will find it usable. There is a second question now: can the agents acting for those customers operate it reliably?

The practical stakes are already significant. An agent working on a platform with good AX finishes tasks in fewer steps and fewer retries, which means lower compute costs and less latency for the end user. It is also much less likely to take the wrong action when it can clearly see its options, their consequences, and the current state of a process. Bad AX is inefficient, and it also creates openings for error.

Early research points the same way. In a July 2026 preprint by Said Elnaffar and Farzad Rashidi, an agent-ready version of an e-commerce prototype was completed successfully in 89.3% of test runs, against 49.3% for an otherwise identical baseline, and agents needed 6.49 steps on average instead of 9.31. The study is small, covering one prototype site, five tasks, and three models, so treat it as a directional signal rather than a benchmark.

Trust suffers quietly. When an agent fails to complete a task, users rarely stop to ask whether the underlying site had ambiguous controls or hidden state. They just see an assistant that failed, and poor AX damages how reliable agentic experiences feel overall.

Reach is at stake as well. As agents increasingly mediate discovery, comparison, and transactions, being legible to them is becoming another dimension of digital discoverability and operability. A platform that humans can use but agents consistently struggle with risks becoming less useful in agent-mediated journeys, which is why AX belongs in the design and engineering conversation from day one rather than in a separate AI initiative.

A starting point for designing with AX in mind

Designing for agents doesn’t mean building two separate products, one for humans and one for machines. In most cases, good AX comes from making the existing product more explicit, structured, and predictable.

• Label actions by intent. Controls should say what they do, not just where they sit in a sequence. “Continue to payment” tells both humans and agents more than “Continue,” and “Confirm booking” is clearer than “Submit.” The consequence of an action should be clear before it’s taken.
• Make state visible. Progress through a flow, applied discounts, availability, validation errors, changing prices, selected options, and transaction status should all be explicitly represented. Agents shouldn’t have to reconstruct application state from visual clues.
• Structure content for parsing. Comparison tables, pricing grids, contract terms, product specs, and order summaries need real semantic structure. If information matters to a decision, it shouldn’t exist only as visual decoration.
• Make errors actionable. An agent needs more than “Something went wrong.” A useful error explains what failed, why when possible, and whether the action can be retried or corrected. This is already a core principle of good DX, and AX extends it to interactions that happen through the product itself.
• Design actions to be safe and predictable. Once agents can perform actions and not just retrieve information, predictability becomes critical. Operations with side effects should behave consistently, repeated actions shouldn’t create duplicate transactions by accident, and high-impact or irreversible actions should make their consequences explicit before they run. The more autonomy an agent has, the more these safeguards matter.
• Document critical flows like contracts. Good DX documents APIs. Good AX should bring the same thinking to important user journeys: what actions are available, what information they require, what changes after each one, which states are terminal, which errors can be recovered from, and which actions need explicit confirmation. Treating flows as contracts makes interfaces more predictable for agents, and usually clearer for humans too.

Designing for three audiences

UX made us better designers and DX made us better engineers. AX asks us to bring both together with a new fact in mind: our digital products no longer have just one kind of visitor.

AI agents are autonomous, programmatic clients acting on behalf of real people, so they fit neither the power user mold nor the traditional API mold. Designing for them is ultimately about making sure the people who delegate tasks to those machines get accurate and predictable outcomes they can trust.

Teams that build AX into their process from the start end up with platforms that are easier for assistants to navigate, cheaper to operate against, and less prone to agent-driven errors. Many of the same principles that make a platform agent-friendly also make it a better platform for humans and developers: explicit intent, semantic structure, visible state, predictable actions, and useful errors.

Agents are already showing up on your platform, so the question is whether it is ready for them.

Note on the data

The figures in this article are snapshots of a fast-changing ecosystem, not permanent baselines. HUMAN Security reported Comet, Atlas, and Claude as the three largest sources of observed agentic traffic in April 2026, and by June Claude had moved ahead of Atlas. Media led e-commerce as the largest destination in April; by June, e-commerce had moved ahead.

The rankings will keep changing. What matters more is the broader pattern: growing agentic activity across discovery, commerce, authentication, accounts, and transactional flows. Check the latest Cloudflare Radar and HUMAN Security data before reusing these figures in future publications.